rework the final-scale and make it possible to have a sharpen after downscaling

[TurboGit]

For the final-scale there is discussion to happen here.

For the final sharpen there is also discussion to happen here.

I had this proposal:

• Introduce a new hidden module named finalsharpen
• Make this module disabled by default
• This new module has a single parameter strength to allow for setting the amount of sharpening (0 = None)
• A new slider "final sharpening" is introduced into the export module (0 by default) up to (100%)
• When this slider is > 0 during export the finalsharpen module is enabled
• The actual strength is passed to it using config key

This issue is to track down the discussion for the implementation of those two issues. Please keep this discussion focused.

EDIT: Initial discussion was in #13635.

EDIT: From reference, having the finalscale doing the job in linear data is a good goal for quality.

EDIT: Moving downscale before any tone curve (Sigmoid, Filmic...) means that all modules after that will work in down-scaled images, what is the impact on quality?

[wpferguson]

What kind of sharpening will it be?

If we use unsharp masking the a threshold could be introduced to avoid sharpening noise.

I implemented sharpening after export with postsharpen.lua. I exported the image, then ran it through Imagemagick's convert usually using the unsharp sharpening. I also tried just using the sharpen function of convert, but got better results using unsharp because of the noise threshold. One thing I did notice was larger file sizes after sharpening.

Marco Carrarini implemented Richardson-Lucy output sharpening in RL_out_sharp.lua. He used GMIC to do the sharpening. Since we link the gmic library, you might be able to use some of the sharpening routines in there. For that matter one of the magick libraries is linked in so that should be available too.

I've pretty much stopped post sharpening. Too often the images were (almost) over detailed, especially after the introduction of Diffuse or Sharpen and the move to a higher resolution camera.

[Mark-64]

Marco Carrarini implemented Richardson-Lucy output sharpening in RL_out_sharp.lua

Yes, RL algorithm does a really nice job and it doesn't amplify noise.

I've pretty much stopped post sharpening. Too often the images were (almost) over detailed, especially after the introduction of Diffuse or Sharpen and the move to a higher resolution camera.

I also hate oversharpened images and my D&S preset is very mild, but I still find useful to apply some small output sharpening after downscaling

[jenshannoschwalm]

Just some first comments from my side, if you want to understand how the pixelpipe works a reminder: -d pipe will give you all the required information. Also if you want to analyse in depth the processed data of a module you can use --dump-pipe a,b,c with ab and c being any module like demosaic, gamma or whatever you are interested in.

Some points we have to remember:

1. Whenever we work on an image in darkroom we use interpolated data depending on the part of the image presented in darkroom (ROI), this interpolation (crop & scale) is done either as part of demosaic for raws or as the first part of the pixelpipe for non-raws.
2. While doing exports the above description is mostly correct if in non-hq mode, with high quality we process full data (no ROI) and apply finalscale at the end of pixelpipe.
3. The interpolator is chosen via UI preference, default is 'lanczos3' tending to overshoot but gives some sharpness.
4. gamma is not used while exporting.

So

1. As hq export and develop in darkroom differ on scaled/unscaled data there can be a significant problem depending on the used interpolator (moire effects, ...) especially on synthetic images. Easy to reproduce in darkroom with the test image provided in sRGB export wrong without linear scaling calculation #13335
2. Point 1 explains for me why just putting finalscale to another position in the pipeline doesn't solve that problem at all. It seems to be not a good "solution"
3. The "bad color" discussed in sRGB export wrong without linear scaling calculation #13335 is mostly not related to non-linear data processed
4. We have a lot of modules now that process data using current scale into account. That leads to sometimes non-avoidable differences ...

How to proceed? Not sure, some ideas

1. If we want some sharpening for exports, i think @TurboGit 's suggestion of a 'finalsharpen' for high-quality exports is a good idea. (Also could be done in finalscale itself though)
2. remove high-quality from export options and always do so for true exports (not for thumbs) a) for quality and b) more stable results
3. can we have a better interpolator for pixelpipe scaling?
4. I wouldn't be in favour of exposing finalscale to be moved around.